Dvidia/piper-action-capture-v01

DVIDIA PiPER action capture v0.1

A capture engineering milestone: actual rendered RGB-D, measured native encoders and authored actuator commands. Two 20-second runs and one interrupted prefix retain 2,005 snapshots and 4,010 RGB/depth files. The two full runs repeat one fixed start/controller/seed; they are not independent examples or a generalized success rate.

Inspect the recorded capture · Research · Source preview release

Download piper-action-capture-v01.zip, extract it and run python -m http.server 8000 for the complete local PNG/NPY report. The small online viewer uses lossy RGB-derived MP4; original arrays, every motor label and independent audits remain in the ZIP. Native task flags remain separate from the recordings' unverified/unknown task results. Four object-tracker channels remain missing/null, so strict learning-target export is refused. CGL depth precision warnings remain explicit; exact repetition is not depth accuracy.

The included wheel is source preview 0.1.3.dev1, distinct from stable data-only SDK 0.1.2. Its isolated installed smoke collected five native intervals/twelve frame files using existing pinned assets. Optional native dependencies and a working OpenGL context are required for capture; no minimum hardware or hard real-time claim. Pillow 12.0.0 is additionally needed only to reproduce the independent image-decoding audit.

DVIDIA source/evidence and PiPER render derivatives: MIT. The preview wheel includes MuJoCo/WASM and other dependencies retaining Apache-2.0, MIT and their original notices; the card's MIT label does not relicense them. No operator corrections, camera pose tracker, trained skill, human-video action recovery or physical robot qualification. Original development refusal/smoke evidence is separately retained, outside the three registered cases.

DVIDIA research · @iammrriver · hello@dvidia.org.